openvinotoolkit/openvino.genai
May 2026 – May 2026
1 Month active
Languages Used
Python
Technical Skills
Deep LearningMachine LearningModel OptimizationPyTorch
Artem Vasenin
PROFILE
Artem Vasenin
Worked on the openvinotoolkit/openvino.genai repository to address a critical runtime bug in the GPT-OSS-20B model, focusing on stabilizing evaluation workflows. Using Python and PyTorch, implemented an automatic tensor data type selection mechanism to resolve mismatches between BF16 and float during model gating and forward passes. This approach included a dtype-auto workaround in model loaders, ensuring robust handling of MXFP4-quantized weights and compatibility with Hugging Face Transformers. The solution was validated through comprehensive end-to-end benchmarking and conformance checks, resulting in reduced runtime errors, faster benchmarking cycles, and improved production readiness for deep learning and model optimization workloads.
Overall Statistics
Feature vs Bugs
0%Features
Repository Contributions
1Total
Bugs
1
Commits
1
Features
0
Lines of code
20
Activity Months1
Work History
May 2026
1 Commits
May 1, 2026OpenVINO GenAI - 2026-05: Delivered a critical runtime bug fix for GPT-OSS-20B by enabling automatic tensor dtype selection (torch_dtype='auto'), addressing a runtime mismatch between BF16 and float during gating/forward passes. The change stabilizes the end-to-end evaluation in openvino.genai bench and improves portability across environments. Related work included introducing a dtype-auto workaround in model_loaders to ensure consistent dtype handling across MXFP4-quantized weights and HF Transformers. Verified via an end-to-end WhowhatBench run and conformance checks. Co-authored PR; pre-commit checks passed. Business impact: reduced runtime errors, faster benchmarking iterations, and stronger production readiness for GPT-OSS-20B workloads.
1 Commits
May 1, 2026OpenVINO GenAI - 2026-05: Delivered a critical runtime bug fix for GPT-OSS-20B by enabling automatic tensor dtype selection (torch_dtype='auto'), addressing a runtime mismatch between BF16 and float during gating/forward passes. The change stabilizes the end-to-end evaluation in openvino.genai bench and improves portability across environments. Related work included introducing a dtype-auto workaround in model_loaders to ensure consistent dtype handling across MXFP4-quantized weights and HF Transformers. Verified via an end-to-end WhowhatBench run and conformance checks. Co-authored PR; pre-commit checks passed. Business impact: reduced runtime errors, faster benchmarking iterations, and stronger production readiness for GPT-OSS-20B workloads.
May 2026
Activity
Loading activity data...
Quality Metrics
Correctness100.0%
Maintainability80.0%
Architecture80.0%
Performance80.0%
AI Usage40.0%
Skills & Technologies
Programming Languages
Python
Technical Skills
Deep LearningMachine LearningModel OptimizationPyTorch
Repositories Contributed To
Overview of all repositories you've contributed to across your timeline